NYC 311 Service Demand Forecasting

Stakeholder forecast dashboard • fixed benchmark window 2022-01-01 through 2026-09-30 • 28-day planning horizon

Forecast demand

10,842

requests per day

Preferred model WAPE

7.3%

28-day rolling mean

Illustrative staffing

284

FTE at point forecast

Category median WAPE

28.8%

8 high-volume complaint types

Planning uncertainty

Point 10,842 requests/day

80% planning band 9,528–12,156 318 FTE at upper band

95% planning band 8,864–12,820 336 FTE at upper band

Bands are empirical planning ranges based on rolling-origin backtest errors, not parametric confidence intervals.

Operational interpretation

Citywide daily demand is comparatively stable: the simple 28-day rolling mean outperformed the more complex statistical and machine-learning benchmarks on mean out-of-time WAPE. Category-level demand is much less uniform, so service lines should be forecast using models matched to their behavior.

Staffing assumes 45 requests per FTE per day at 85.0% target utilization. This is an illustrative scenario, not an NYC staffing standard.

Citywide model comparison

Model	Mean WAPE	Relative error
28-day rolling mean	7.3%
Weekly ETS	7.3%
HistGradientBoosting	7.7%
Naive	8.1%
7-day seasonal naive	9.1%

Category forecastability

The same 28-day rolling-mean benchmark is applied to the eight highest-volume complaint types. Lower WAPE indicates better forecast accuracy.

Complaint type	Requests	Mean WAPE
Illegal Parking	2,434,412	8.6%
Blocked Driveway	812,009	9.1%
Abandoned Vehicle	322,172	20.5%
UNSANITARY CONDITION	559,238	27.3%
Noise - Residential	1,800,082	30.2%
Street Condition	391,653	34.1%
Noise - Street/Sidewalk	780,567	47.9%
HEAT/HOT WATER	1,289,976	53.0%

Targeted model improvements

Volatile categories were retested with annual persistence and calendar/lag machine learning rather than assuming the citywide model was optimal.

Complaint type	Rolling mean	Best model	Best WAPE	Improvement
HEAT/HOT WATER	53.0%	Annual naive (365d)	35.2%	33.5%
Noise - Street/Sidewalk	47.9%	Calendar/lag HGB	30.7%	35.9%
Street Condition	34.1%	Calendar/lag HGB	23.7%	30.4%

Decision takeaway

Use simple models where they work. Aggregate 311 demand is forecastable with a recent-level baseline, while volatile categories benefit from service-specific seasonal or lag/calendar models. Forecast uncertainty should be translated into planning ranges rather than a single staffing number.

Source: NYC Open Data, 311 Service Requests from 2020 to Present (dataset erm2-nwe9). Forecast benchmark uses a fixed historical analysis window for reproducibility.

View repository · Research design · Metrics glossary